Publication 02 · Aspect-level sentiment

Bidirectional Directed Acyclic Graph Neural Network for ASC.

Rebuilding the dependency tree as a bidirectional directed acyclic graph — so multi-hop sentiment information can actually travel.

Published · TALLIP 2025 Aspect-level sentiment Directed acyclic graph SenticNet knowledge

Junjie Xu*, Luwei Xiao*, Anran Wu, Tianlong Ma, Daoguo Dong, Liang He · East China Normal University (* equal contribution)

GNN-over-dependency-tree methods shorten the distance between aspect terms and opinion words — until depth causes over-smoothing and long-distance relations fade. BDAGNN rebuilds the syntax as a bidirectional directed acyclic graph, propagates sentiment information multi-hop without stacking ever-deeper layers, and injects affective commonsense from SenticNet — reaching partial state-of-the-art on four benchmarks.

4benchmark datasets evaluated
×2propagation directions in the reconstructed DAG
Multi-hopsentiment information flow without deeper stacks
SenticNetaffective commonsense injected into the graph
01

The question

Aspect-level sentiment classification asks a local question with global context: is this aspect positive, negative, or neutral in this sentence? Syntax helps — a dependency tree links "food" to "awful" in two hops. But standard GNN encoders pay for depth with over-smoothing: stack layers to reach distant opinion words, and node representations blur together.

The question: can the syntactic structure itself be rebuilt so that long-distance sentiment evidence survives propagation?

02

The approach

BDAGNN makes two structural moves. First, the bidirectional directed acyclic graph (BDAG) reconstructs the dependency tree — replacing a single rooted view with coordinated bottom-up and top-down directions over an acyclic structure. Second, the BDAGNN encoder propagates sentiment information multi-hop through this structure efficiently, instead of relying on many stacked GNN layers.

Affective commonsense from SenticNet enriches the graph's edges, so classification can consult world knowledge about sentiment, not only sentence-internal co-occurrence.

BDAGNN framework: word embedding, dependency tree, bidirectional DAGs, multi-hop information transmission, and aspect-specific masking
Fig. 1The BDAGNN framework — from dependency tree to bidirectional DAGs, through multi-hop information transmission, to aspect-specific masking and prediction.
03

What it shows

BDAGNN obtains partial state-of-the-art performance across four benchmark datasets — comparable to, and on several settings better than, the strongest GNN-based baselines.

Ablations vary the number of BDAG layers and directions: accuracy and F1 respond as the structure predicts, supporting the claim that it is the bidirectional acyclic reconstruction — not added capacity alone — that carries the improvement.

Accuracy and F1 of BDAGNN + BERT with different numbers of BDAG layers
Fig. 2Accuracy (left) and F1 (right) as BDAG depth varies — the effect of the structure is measurable, not incidental.
04

Claim boundary

✓ Established

  • A bidirectional DAG reconstruction of syntactic dependencies
  • Multi-hop sentiment propagation that avoids over-smoothing by depth-stacking
  • SenticNet affective knowledge integrated into the graph
  • Partial state-of-the-art on four benchmarks, with ablations and analysis

✗ Explicitly not claimed

  • "Partial" is stated deliberately — BDAGNN does not dominate every setting
  • English benchmark sentences; no cross-lingual claim
  • No claim that the DAG mirrors human syntactic processing
  • Gains are model-side; no human-subject evaluation
05

Citation

@article{xu2025bdagnn,
  title   = {Bidirectional Directed Acyclic Graph Neural Network
             for Aspect-level Sentiment Classification},
  author  = {Xu, Junjie and Xiao, Luwei and Wu, Anran and Ma, Tianlong
             and Dong, Daoguo and He, Liang},
  journal = {ACM Transactions on Asian and Low-Resource Language
             Information Processing},
  volume  = {24},
  number  = {4},
  pages   = {33:1--33:14},
  year    = {2025},
  doi     = {10.1145/3716501}
}